DVFL: A Lightweight Verifiable Federated Learning System With Dynamic Obfuscation for Resource-Constrained IoT Devices

The proliferation of IoT devices necessitates privacy-preserving on-device training, yet existing Federated Learning systems struggle with privacy leakage and unverifiable aggregation under resource constraints. In this paper, we present DVFL, a lightweight verifiable FL system with dynamic obfuscation tailored for IoT. Distinct from traditional FL, DVFL adopts a decoupled personalization strategy: it trains a globally shared feature extractor while keeping the privacy-sensitive classification head strictly local, replacing it with a normalized dummy value during uplink. This design, supported by a 2-of-2 helper module for zero-sum masking, inherently prevents leakage from sensitive layers while enabling personalized adaptation. Furthermore, we introduce a lightweight dual-verification mechanism that allows clients to independently validate aggregation integrity using low-cost hash operations, avoiding prohibitive digital signatures. Experimental results on MNIST, CIFAR-10, UCI HAR and MIT BIH show that DVFL reduces Membership Inference Attack AUC to ≈ 50% and thwarts gradient leakage, while achieving personalized accuracy comparable to undefended baselines.